Artificial intelligence in adrenal imaging: A critical review of current applications

Adrenal lesions (AL) are found on 3 to 7% of routine abdominal imaging examinations and most of them are benign and asymptomatic and require no further treatment [1,2]. However, AL encompasses a wide spectrum of conditions, and definite characterization is needed to determine the best management. In this regard, in adults, more than 75% of incidentally found ALs (i.e., incidentalomas) are adrenal adenomas (AAs) without hormone hypersecretion [2]. However, 1–5% of adrenal incidentalomas are pheochromocytomas, 1–12% are secreting adenomas, and 8% are malignant ALs [2,3]. Current characterization of AL with imaging is based on the use of quantitative and qualitative criteria [4,5]. However, these criteria have limited capabilities for a definite characterization of atypical AL [4,5].

Radiomics refers to a complex process based on myriad features extraction from medical imaging using quantitative methods that mainly provide estimates of image heterogeneity [6]. Extracted features include first order features (e.g., histogram and shape) and second order features that consider spatial location and interrelationship between voxels and require a matrixial transformation [7,8]. For each matrix considered, similar features related to heterogeneity may be extracted and ultimately, thousands of parameters can be extracted for further analysis [8]. The grey level co-occurrence matrix is one of the most used matrix for transformation [8]. The ultimate goal of radiomics is to extract biomarkers predictive of survival or response to treatments on pre-operative imaging like those that can be obtained invasively with transcriptomics or genomics [9]. Radiomics allows developing quantitative scores to differentiate benign from malignant AL, identify histological subtypes, or develop some prognostic biomarkers when validated [10], [11], [12], [13].

Artificial intelligence (AI) is a general term that refers to the ability of a computer algorithm to imitate human brain and perform tasks such as learning and problem solving [14,15]. AI includes machine learning (ML) and deep learning (DL) [9]. Briefly, ML consists in providing more data to continuously update the computer/software performance for a given task. This implies that the computer/software learns from the data and improves its performances through experience [14]. DL, a subset of ML, refers to the use of a layered structure of algorithms (i.e., the deep neural network), which is aimed to imitate the human brain perception in recognizing features from input data [14]. A complete AI algorithm includes detection, segmentation, analysis, and classification [16] (Fig. 1). However, some steps such as detection and segmentation may be done by humans, and to date, most AI studies only focused on only one step of the process [16].

Radiomics can be associated with AI, and these joint approaches are increasingly being developed. Indeed, radiomic provides a large amount of data, and AI allows to process them better than conventional statistical methods. Reciprocally, AI models are based on the metrics provided by radiomic approaches. In the field of medical imaging, AI suffers from many biases induced by a lack of standardization and methodological flaw [17]. To avoid these biases responsible of many limitations in evidence-based medicine and daily practice application, guidelines and quality scores have been developed but they are insufficiently applied [17,18].

The purpose of this article was to sum-up recent developments and current results of AI in the field of adrenal imaging and discuss future perspectives.

留言 (0)

沒有登入
gif